Adaptive Attention Convolutional Neural Network for Liver Tumor Segmentation.

Area of Science:

• Medical Image Processing
• Artificial Intelligence in Healthcare
• Radiotherapy Planning

Background:

• Accurate liver and liver tumor segmentation is crucial for effective radiotherapy.
• Challenges include complex tumor characteristics and low contrast, complicating manual segmentation.
• Existing methods struggle with precise boundary definition and varied tumor appearances.

Purpose of the Study:

• To develop an advanced neural network (S-Net) for automated, end-to-end liver tumor segmentation from CT images.
• To enhance segmentation accuracy by incorporating attention mechanisms and long-hop connections.
• To address limitations of current segmentation techniques for challenging liver tumors.

Main Methods:

• A classical coding-decoding neural network structure was employed for end-to-end segmentation.
• An attention mechanism was integrated to capture long-range semantic information and inter-channel relationships.
• Long-hop connections facilitated the fusion of semantic information from different network paths.
• Closed operations were utilized for noise reduction and removal of small interruptions.

Main Results:

• The S-Net architecture was evaluated on diverse datasets (LiTS, 3DIRCADb, Hubei Cancer Hospital).
• Key performance metrics included Dice Global (DG), Dice per Case (DC), VOE, ASSD, and RMSE.
• Achieved a tumor DG of 0.7555 and DC of 0.613, demonstrating robust segmentation capabilities.
• Performance varied by tumor size, with higher accuracy for larger tumors (DG=0.7819) compared to smaller ones (DG=0.3246).

Conclusions:

• The S-Net, enhanced with attention and long-hop connections, effectively extracts more semantic information.
• Experimental results confirm improved tumor recognition in CT images.
• The proposed method shows significant potential for assisting clinicians in liver tumor diagnosis and treatment planning.